Skip to content
Back to search results

Interrupted Time Series for Public Health

Introduced by
Anita K. Wagner, Stephen B. Soumerai et al. (segmented-regression formulation); James Lopez Bernal, Steven Cummins & Antonio Gasparrini (public-health tutorial)
Year
2002
Type
Quasi-experimental design estimating level and slope changes in a population outcome after an intervention
Family
Quasi-experimental time-series evaluation / public-health policy analysis
Font size

What does it do?

Interrupted time series analysis, usually implemented as segmented regression, is a strong quasi-experimental design for evaluating the effect of a public-health intervention introduced at a known point in time.

By tracking a population-level outcome — prescribing rates, infections, injuries, hospital admissions — over many equally spaced periods before and after the intervention, it asks whether the outcome's level jumped and whether its underlying trend changed when the intervention took effect, relative to the pre-intervention trajectory projected forward as the counterfactual. The segmented-regression formulation was popularized for intervention research by Wagner, Soumerai and colleagues, and Lopez Bernal, Cummins and Gasparrini's 2017 International Journal of Epidemiology tutorial is the standard modern guide for public-health applications, covering autocorrelation, seasonality, and the use of comparison series.

The intuition behind the method

Suppose a city bans smoking in restaurants and you want to know whether heart-attack admissions fell. You cannot run a trial, but you do have monthly admission counts for years before and after the ban. The idea of interrupted time series is simple: draw the trend line through the before-period data and extend it forward as your best guess of what would have happened without the ban — that projection is the counterfactual. Then look at what actually happened after the ban and measure two things. First, did the line jump up or down right at the moment of the ban — a sudden level change? Second, did the line's slope tilt — did admissions start climbing or falling faster or slower than before? Together these capture an immediate effect and a gradual one. The honest version of this analysis worries about a few things that can fool you: outcomes in adjacent months are correlated, there are seasonal swings (flu season, summer), and other events may coincide with the intervention. So good practice models the autocorrelation and seasonality and, whenever possible, compares against a similar population that did not get the intervention to rule out coincident trends.

History

Segmented regression of time-series data has roots in quasi-experimental design and was adapted for evaluating policy and clinical interventions in the health-services literature. Wagner, Soumerai, Zhang and Ross-Degnan's 2002 paper in the Journal of Clinical Pharmacy and Therapeutics gave the approach a clear, accessible formulation for medication-use and policy research, defining the level and slope-change parameters that practitioners now report. As routinely collected public-health data proliferated, the method became a leading tool for evaluating interventions that cannot be randomized. Lopez Bernal, Cummins and Gasparrini's 2017 International Journal of Epidemiology tutorial consolidated best practice for public-health applications, with explicit guidance on autocorrelation, seasonality, model checking, comparison series, and counterfactual prediction, and it has become the standard reference, complementing related designs such as synthetic control for policy evaluation.

When to use it

Use interrupted time series for public health when an intervention or policy was introduced at a clearly identifiable time and you have a reasonably long run of equally spaced outcome measurements before and after it — the more pre-intervention points the better, since the design leans on a stable pre-trend. It is well suited to evaluating smoking bans, prescribing-guideline changes, vaccination campaigns, traffic or alcohol laws, and similar population-level interventions where randomization is impossible and the outcome is measured routinely. It is strongest when a comparison series is available and when the intervention's timing is sharp rather than gradual or anticipated. It is a poor choice when the series is short, when the intervention coincides with other major events that cannot be separated, when the outcome is measured only a few times, or when the effect is expected to be so gradual or delayed that distinguishing it from the underlying trend is hopeless. For single-unit policies with rich donor pools, synthetic control is a complementary or alternative design.

Limitations and considerations

  • Requires a sufficiently long, stable pre-intervention series; short series give unreliable trends and weak power.
  • Vulnerable to confounding by other events occurring at or near the intervention time unless a control series is used.
  • Autocorrelation and seasonality must be modeled correctly or significance is overstated and effects misattributed.
  • Extrapolating the pre-trend far into the post period can be fragile, especially for slow-onset or anticipated effects.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Estimates both an immediate level change and a change in trend, capturing distinct kinds of intervention effect.
  • Uses the pre-intervention trend as an explicit, projectable counterfactual, making assumptions transparent and visual.
  • Strong quasi-experimental design for routinely collected population data where randomization is impossible.
  • Greatly strengthened by adding a comparison series, which controls for events coinciding with the intervention.

Applications

  • Evaluating whether a public smoking ban reduced hospital admissions for cardiovascular or respiratory disease.
  • Assessing the impact of a prescribing guideline, formulary change, or drug-safety warning on medication use.
  • Measuring the effect of a vaccination campaign or screening program on disease incidence.
  • Studying traffic-safety or alcohol laws on injuries, crashes, or alcohol-related harm.
  • Evaluating the effect of an emergency public-health measure on infections, mortality, or service use during an outbreak.

How it works

How it works

1

The analysis begins by specifying a segmented regression of the outcome on time. The model includes a continuous time variable capturing the pre-intervention trend, an intervention indicator that switches on at the intervention point, and an interaction between time and that indicator that allows the slope to change afterward. This single equation encodes the two segments — before and after — sharing the same time axis. Wagner and colleagues set out this formulation as the workhorse of intervention time-series research, with each coefficient mapping onto an interpretable feature of the before-and-after trajectory.

Formula
Y_t=\beta_0+\beta_1\,t+\beta_2\,D_t+\beta_3\,(t-T_0)D_t+\varepsilon_t,\quad D_t=\mathbb{1}[t> T_0]
2

From the fitted model, two coefficients carry the causal interpretation. The coefficient on the intervention indicator is the immediate change in the level of the outcome at the moment of the intervention — a step up or down. The coefficient on the post-intervention time interaction is the change in slope, that is, how much the trend tilts after the intervention compared with before. Distinguishing these is central: a policy may produce an abrupt one-off shift, a gradual change in trajectory, both, or neither, and Wagner and colleagues emphasize reporting both effects rather than a single before-after difference.

Formula
\beta_2=\text{immediate level change},\quad \beta_3=\text{change in trend (slope)}
3

Time-series outcomes violate the independence assumption of ordinary regression because observations close in time are correlated, and many public-health series have strong seasonal cycles. The analyst must therefore model the error structure — for example with autoregressive terms or Newey-West-type corrections — and include seasonal terms such as harmonic (sine and cosine) functions or month indicators. Lopez Bernal, Cummins and Gasparrini stress that ignoring autocorrelation produces standard errors that are too small and overstated significance, and that unmodeled seasonality can masquerade as an intervention effect, so both must be handled explicitly.

Formula
\varepsilon_t=\rho\,\varepsilon_{t-1}+u_t;\quad +\textstyle\sum_k \bigl[\gamma_k\sin(\tfrac{2\pi k t}{s})+\delta_k\cos(\tfrac{2\pi k t}{s})\bigr]
4

Whenever possible, the design is strengthened by adding a control series: an equivalent outcome from a population or area not exposed to the intervention. Interacting the group indicator with the intervention and post-intervention time terms turns the analysis into a controlled interrupted time series, where the effect is the difference in level and slope changes between the treated and control series. This guards against confounding by events that coincide with the intervention but affect both groups, and Lopez Bernal and colleagues present the comparison series as one of the most powerful ways to bolster causal claims from ITS.

Formula
Y_t=\dots+\alpha\,G+\theta_2\,(G\cdot D_t)+\theta_3\,(G\cdot(t-T_0)D_t),\quad G\in\{0,1\}
5

Finally, the intervention's effect is quantified by contrasting the observed post-intervention outcomes with the counterfactual obtained by projecting the pre-intervention trend forward as if nothing had changed. The vertical distance between the actual series and this extrapolated trend, at each post-intervention time point, is the estimated effect, which can be summarized as an absolute or relative difference at chosen horizons. Lopez Bernal and colleagues recommend plotting the observed series, the fitted segments, and the counterfactual together, both to communicate the result and to expose any poor fit or implausible extrapolation.

Formula
\widehat{\text{effect}}_t=Y_t-\hat{Y}_t^{\,\text{counterfactual}},\quad \hat{Y}_t^{\,cf}=\hat{\beta}_0+\hat{\beta}_1 t

Common pitfalls

  • Fitting ordinary least squares without correcting for autocorrelation, which deflates standard errors.
  • Failing to model seasonality, so seasonal swings are mistaken for an intervention effect.
  • Using too few pre-intervention points to establish a credible baseline trend.
  • Attributing the post-intervention change to the policy when a coincident event affecting the same population is the real cause and no comparison series is included.

Frequently asked questions

What is the difference between the level change and the slope change?

The level change is the immediate, step-like shift in the outcome at the moment the intervention takes effect — the jump up or down right at the interruption. The slope change is how much the underlying trend tilts afterward, that is, whether the outcome starts rising or falling faster or slower than it did before. A policy might cause an abrupt drop with no change in trend, a gradual change in trend with no immediate jump, both, or neither. Wagner and colleagues stress reporting both because they describe fundamentally different effect patterns.

Why do I need to worry about autocorrelation?

In a time series, observations close together are correlated, so the residuals of an ordinary regression are not independent. Ignoring this leaves the point estimates roughly unbiased but makes standard errors too small, inflating significance and risking false positives. Lopez Bernal, Cummins and Gasparrini recommend testing for autocorrelation (for example with the Durbin-Watson statistic or autocorrelation plots) and accounting for it through autoregressive error models or robust standard errors. Seasonality should be modeled at the same time, since seasonal cycles otherwise contaminate both the trend and the estimated effect.

How does a comparison series strengthen the design?

A single interrupted series can be confounded by anything else that happened around the intervention time. Adding an equivalent series from a population that did not receive the intervention turns the analysis into a controlled interrupted time series: the effect becomes the difference between the treated and control series in their level and slope changes. If both series are subject to the same coincident shocks but only the treated one received the intervention, the comparison removes those shared influences. Lopez Bernal and colleagues regard a well-chosen comparison series as one of the most effective ways to sharpen causal inference from ITS.

Related methods

Sources

  1. Wagner, A. K., Soumerai, S. B., Zhang, F., & Ross-Degnan, D. (2002). Segmented Regression Analysis of Interrupted Time Series Studies in Medication Use Research. Journal of Clinical Pharmacy and Therapeutics, 27(4), 299-309.

    Open DOI reference
    Copied.
  2. Lopez Bernal, J., Cummins, S., & Gasparrini, A. (2017). Interrupted Time Series Regression for the Evaluation of Public Health Interventions: A Tutorial. International Journal of Epidemiology, 46(1), 348-355.

    Open DOI reference
    Copied.